--- In IBObjects@y..., "Jason Wharton" <jwharton@i...> wrote:

> This question and the one after it are both related to what I want

to say.

>
> The transaction OAT processing is causing these fetches. I am not

quite done

> fully putting in place all of the nice things horizontal dataset

refinement

> will allow. All this extra fetching will go away on these datasets

with the

> benefits of this new feature because there is no point in the

fetches with

> the ability to refine a dataset. It is non refined datasets that

need to

> exhaust the cursor to allow the transaction to be released. With

refinement

> I can close and reopen a cursor anywhere in the entire dataset.

That is the

> whole point of the feature to gain an added level of freedom from

holding a

> cursor open.
>
> I'll explain it a bit better if I can. The new refinement is able

to open a

> cursor and walk a few records and then if necessary, like when

Locate or

> Last is called it kills the current cursor and generates a new one

where the

> records of interest are. There is no point in trying to exhaust

records out

> of a cursor when it can kill itself and regenerate itself anywhere

in the

> dataset required.
>
> So, what you should look forward to (hopefully before I make the

final

> release) is this background fetching going away. Cursors simply

need to be

> truncated (killed) so that OAT processing will go on totally

unobstructed.

> That is a totally cheap operation under the architecture of the

refinement

> process. The only inefficiency it will impose is records that have

been

> batched together in network packets may have to be requested again

if a new

> boundary to the current cursor is defined.
>
> Perhaps you have noticed performance when doing individual fetches.

10 or so

> will go by quickly and then you hit one fetch that takes just a

pause. Each

> time records are sent, InterBase prefetches and batches them

together. If it

> gets half way through a batch and truncates the cursor it will

fetch records

> from the already prefetched batch. Not a big deal if this only

happens on a

> basis of once in 30 seconds at most. Depends on how you have tuned

the

> transaction timeout parameters.
>
> Hope this makes sense. If not, don't worry about it. I've got it all
> (fairly) under control.

What should I do?